#!/usr/bin/env bash
# Direct Q4 probe: AWQ INT4 language weights plus FP4 KV.  It refuses to
# continue if GPU1 allocation exceeds the user's 17,068 MiB llama ceiling.
set -euo pipefail

ROOT=/home/lab/bench/ornith/modular
OUT="$ROOT/sglang/results/q4-cap17"
CAP_MIB=17068
mkdir -p "$OUT"

systemctl --user stop ornith-sglang-q4cap 2>/dev/null || true
systemctl --user reset-failed ornith-sglang-q4cap 2>/dev/null || true
systemd-run --user --unit=ornith-sglang-q4cap --collect \
  -p TimeoutStopSec=30 \
  -p "StandardOutput=append:$OUT/server.log" \
  -p "StandardError=append:$OUT/server.log" \
  --setenv=ORNITH_CONTEXT=32768 \
  --setenv=ORNITH_CONCURRENCY=3 \
  --setenv=ORNITH_MEMORY=0.55 \
  --setenv=ORNITH_GRAPHS=0 \
  /bin/bash "$ROOT/sglang/frozen/experiments-20260918-v1.sh" awq-fp4-eager

ready=0
for _ in $(seq 1 120); do
  if curl -fsS --max-time 2 http://127.0.0.1:8130/health >/dev/null; then ready=1; break; fi
  if ! systemctl --user is-active --quiet ornith-sglang-q4cap; then break; fi
  sleep 2
done

used=$(nvidia-smi -i GPU-f4a988cb-c310-b618-5a9c-4462d4d33add --query-gpu=memory.used --format=csv,noheader,nounits | awk '{print int($1)}')
printf '{"ready":%s,"memory_mib":%s,"cap_mib":%s,"weights":"AWQ_INT4","kv":"FP4"}\n' "$ready" "$used" "$CAP_MIB" | tee "$OUT/measurement.json"
if [ "$ready" -ne 1 ] || [ "$used" -gt "$CAP_MIB" ]; then
  systemctl --user stop ornith-sglang-q4cap || true
  exit 1
fi

/home/lab/venvs/vllm/bin/python "$ROOT/shared/quality.py" \
  --base http://127.0.0.1:8130/v1 --model ornith-sglang --out "$OUT/quality.json"
